Imagine a voice-activated system used to admit patients to a hospital. The command might be "admit patient," but what if "register patient" is easier for the system to recognize, especially in a loud hospital environment?

KWS algorithms must respond to such commands immediately, correctly, despite running on resource-constrained hardware. With speed being of the essence, and network latency unacceptable, the algorithms need to execute locally on the device. Improved security and privacy are another benefit of this approach.

To date, research has prioritized techniques that enable local execution, like model quantization and pruning, delivering impressive gains. But even the most efficient model can struggle if the chosen keywords are hard to detect, especially in noisy spaces. We believe success and failure may depend on the structures of the keywords itself, such as the morphemes of the words.

A morpheme is the smallest meaningful part of a word that has meaning, like “re” in “redo” or “ing” in “running.” Our hypothesis: selecting keywords with less-common morphemes (like “gister” in “register”) will lead to better detection than with common ones. Better yet, can we design an optimized set of keywords for such KWS systems?

Our team used a two-stage approach:

1. Text-Based Analysis
   The goal of this work stream was to examine the entropy of morphemes - how predictable they were within words.

   Using OpenAI’s GPT model, we generated a list of healthcare-specific keywords and their synonyms, along with twenty or more sentences using each word. The LLM then broke each word into its morphemes and assessed each one using entropy metrics:

   • Weighted entropy reflects the rarity of morphemes using a uniqueness scale.
   • Joint entropy measures how rarely those morphemes appear together in a keyword. A higher joint entropy value meant the keyword was uncommon.
   
   

   As source material, we used frequently used dialogue datasets¹, extracting sentences, and segmenting words into morphemes. We then calculated the entropy values for each morpheme we identified based on their frequency.




2. Speech-Based Analysis
   This stage explicitly accounts for physical intelligence factors, including acoustic noise, signal distortion, and sensing limitations, which strongly influence real-world keyword detectability but are often abstracted away in model-only evaluations.

   While entropy provides insight into keyword selection, it is equally important to evaluate how detectable they are in real world scenarios. To do that, we generated speech samples of sentences containing the target keywords and their synonyms. We use this process:

   1. Converted sample sentences into audio using ChatterboxTTS (text-to-speech). Each sample sentence contained the candidate keyword(s).
   2. Added background noise with varying signal-to-noise ratios (SNR) from the PASCAL CHiME dataset to simulate real-world conditions.
   3. Denoised the sample(s) using the DeepFilterNet model.
   4. Next, we used Whisper, an automatic speech recognition model, to convert the denoised audio samples back into sentences.
   5. Finally, we converted both the original sentences and the newly generated sentences into text embeddings (i.e., numerical vectors where similar texts have higher embedding similarity) and compared their similarities. A higher value for a specific keyword indicates its ease of detection in a noisy environment. For better generalization, we report the average similarity across all the sentences for a candidate keyword.
   
   

   This two-stage interdisciplinary pipeline, combining ideas from linguistics and audio research, provides a systematic, reproducible framework for evaluating and selecting keywords. Words that the process detected as similar should inherently be more detectable in noisy environments.

Our study found that words with higher entropy generally have a higher likelihood of being correctly detected by the KWS algorithm. The findings suggest that choosing the right keywords can significantly improve the reliability of voice-controlled systems—especially in noisy environments. For example: "Register patient" outperformed "admit patient", and "deliver medication" surpassed "give drug". These results held up across both text and audio evaluations, with speech similarity scores exceeding 0.9 at -10 dB SNR in the best cases. That is, even in situations where noise levels were almost ten times higher than the speech itself, the keywords identified remained highly effective.

This work provides a practical framework for improving KWS model performance by refining keyword choices. The process can accelerate ongoing model-level use case adaptation and development. By accounting for how language interacts with acoustics, sensing, and environmental noise, keyword selection becomes part of a broader physical intelligence approach at the edge. Future efforts will extend these methods to other industries, including automotive systems, and to languages beyond English. Collaboration with domain experts will further align keyword choices with real-world operating conditions. This alignment helps ensure that voice systems remain accurate, natural, and dependable in the environments where they are deployed.